AIエージェント
agent-lightning
Microsoft's trainer for AI agents - reinforcement learning for almost any agent framework
★ 18.6k スターMIT
ガイドを見る →NeMo-Agent-Toolkitは、NVIDIAが公開しているオープンソースのAIエージェントプロジェクトです（GitHub: NVIDIA/NeMo-Agent-Toolkit）。GitHubスターは2,661、フォークは780で、Apache-2.0ライセンスの下で公開されています。GitHubでの最終更新は2026年10月8日です。
NVIDIA's toolkit for connecting, profiling and optimizing teams of AI agents — across any framework
Decide what you need: build a new workflow, or instrument, profile and evaluate an agent you already have in another framework.
pip install nvidia-nat, adding the extra for your framework; clone the repo if you want the examples.
Write a workflow.yml defining functions (tools), LLMs and the workflow agent, and set your model key (e.g. NVIDIA_API_KEY for NVIDIA NIMs from build.nvidia.com).
Run nat run --config_file workflow.yml --input "..." and inspect the result, then open the built-in UI to chat with and debug the workflow.
Works side-by-side with LangChain, LlamaIndex, CrewAI, Semantic Kernel, Google ADK and custom frameworks to add instrumentation without a rewrite.
Profile entire workflows from agent level down to individual tokens, trace execution and understand runtime behavior in production.
An offline evaluation system, a hyper-parameter and prompt optimizer, and reinforcement-learning fine-tuning support for agent-specific models.
Integrate MCP tools into agents or publish workflows as MCP servers (FastMCP publishing), and build distributed agent teams with the A2A protocol with authentication.
Agent Performance Primitives accelerate graph frameworks (LangChain, CrewAI, Agno) with parallel execution, speculative branching and node-level priority routing.
Build tools and workflows once and reuse them; a built-in chat UI lets you interact with agents and debug workflows visually.
Clone the toolkit if you want the examples locally; the library installs from PyPI in step 2.
git clone https://github.com/NVIDIA/NeMo-Agent-Toolkit.git
cd NeMo-Agent-ToolkitInstall the core package, adding a framework extra such as [langchain] if needed (Python 3.11–3.13).
pip install nvidia-nat
# e.g. pip install "nvidia-nat[langchain]"Set your NVIDIA key, save the README's workflow.yml (a ReAct agent with a Wikipedia tool), and run it with the nat CLI.
export NVIDIA_API_KEY=<your-key>
nat run --config_file workflow.yml --input "List five subspecies of Aardvarks"NVIDIA maintains it, as part of the NeMo ecosystem; docs are at docs.nvidia.com/nemo/agent-toolkit.
Yes, Apache-2.0 licensed. Models you call (NVIDIA NIMs or other providers) have their own pricing.
The Hello World uses NVIDIA NIMs with an NVIDIA_API_KEY from build.nvidia.com; the toolkit itself works with the LLM providers you configure in your workflow.
LangSmith-style tracing is one piece (and the toolkit integrates with LangSmith natively). The toolkit also adds workflow composition, profiling, evaluation, optimization and protocol publishing across frameworks.
pip install nvidia-nat, write a workflow.yml, and run nat run — or open the Colab introduction notebook with no local setup.
Microsoft's trainer for AI agents - reinforcement learning for almost any agent framework
The agent that grows with you — a Python framework for adaptive AI agents.
Policy, identity, and sandboxing guardrails for AI agents.
Swarm — OpenAI's experimental framework for lightweight multi-agent orchestration with agents and handoffs, MIT
Microsoft's GUI agent that grew into UFO³ — orchestrating agents across Windows, Linux and mobile devices.
Koog is a JVM (Java and Kotlin) framework for building predictable, fault-tolerant and enterprise-ready AI agents across all platforms – from backend service…
LibreAIカタログに掲載されたAIエージェントプロジェクトの中で最もスターの多いもの — NeMo-Agent-Toolkitが合わなかった場合に検討したい、現実的な候補リストです。
A personal AI that really does things — on any OS, with your data staying yours.
An agentic skills framework & software development methodology that works.
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
The agent that grows with you — a Python framework for adaptive AI agents.
DeepSeek's plugin-based agent harness where everything is a plugin.
カタログの中でNeMo-Agent-Toolkitとスター数が近いAIエージェントプロジェクト — 前後の棚もそのまま見て回れます。
Bright Data's MCP server: web search, scraping, structured data and browser automation for agents (75 tools)
Rust terminal multiplexer with a typed SDK: drive any CLI app from code.
Self-learning browser infrastructure for AI agents — sites remembered, so agents stop rediscovering the web
The open-source memory layer for autonomous agents — knowledge-graph memory that emulates human memory
Python library for building multimodal language agents — VLM, video and mobile, fast to prototype and ship
Long-term memory runtime for AI agents — plain Markdown as truth, local ranked retrieval, no API key